Odd Lots2026.06.2948 min

How Vertical Integration and Autonomous Agents Drive Commercial AI Success

Original title · Baidu's CFO on How It Became a Full-Stack AI Player

用阿法問答深入這集 ✨

或自己問：

The geopolitical and technical race between US and Chinese artificial intelligence is entering its most practical phase yet: monetizing the application layer. While American hyperscalers capture headlines with massive capital expenditure figures, Chinese tech giants are proving that vertical integration can drive cash-flow-positive growth.

Cloud Platform as the Strategic Anchor

Rather than viewing the AI stack as isolated layers, the priority lies in the cloud platform. Cloud acts as the crucial host that bridges chip infrastructure, proprietary large language models, and external models. For operators, this platform approach is critical because up to 80% of incremental token demand has shifted away from pre-training toward real-time inference.

The Rise of Daily Active Agents

Traditional metrics like daily active users are giving way to daily active agents as the core yardstick of tech performance. Agents are moving beyond simple digital assistants into autonomous systems that plan, execute, and complete complex commercial tasks, such as optimizing logistics at major shipping ports. This shift is redefining corporate purchasing, turning AI from a cost-center IT decision into a top-down CEO priority.

Autonomous Vehicles Shape Commercial AI

Physical AI applications, notably robotaxis, are serving as massive real-world data flywheels. Operating at a scale that closely rivals leading US self-driving platforms, these fleets generate high-quality multimodal training data. The ultimate tipping point for this technology will occur when the operational cost per mile falls below the threshold of car ownership.

Keep a close eye on Baidu's upcoming spin-off and independent listing of its chip unit in Hong Kong.

Assets mentioned in this episode

BIDUBaidu▲ 看多
Baidu has successfully demonstrated a disciplined approach to the AI spending race by maintaining strong cloud revenue growth and positive operating cash flow. The company's vertical integration from custom Kunlun chips to its Ernie foundational model and Apollo Go robotaxis gives it a robust data flywheel. Furthermore, the planned spin-off and separate listing of its chip design asset in Hong Kong should unlock significant shareholder value and allow its hardware to find broader market adoption.
GOOGLAlphabet— 中性
Google continues to lead global AI infrastructure development, leveraging its proprietary TPU architecture to accelerate its Google Cloud Platform growth. Its Waymo self-driving unit remains the premier benchmark for autonomous ride-hailing volumes globally. However, the firm faces rising competition from vertically integrated players in both search capabilities and physical AI execution.

Key questions

Why is the tech industry shifting its focus from pre-training to real-time inference?→

The industry is moving past foundational model building because roughly 80% of incremental token consumption is now driven by real-time inference, necessitating a shift toward operational efficiency and lower-cost, high-speed token delivery.

💡 詢問阿法，看詳細逐字稿分析→

How are companies balancing massive AI infrastructure spending with shareholder returns?→

Firms resolve this 'impossible triangle' through strict lifecycle discipline, ensuring every dollar spent on servers and data centers is evaluated against a 20 to 40-month cash-back cycle to maintain positive operating cash flow.

💡 詢問阿法，看詳細逐字稿分析→

Why are daily active agents replacing daily active users as a key performance metric?→

Unlike human users, autonomous agents can execute complex commercial tasks like logistics and e-commerce. This shifts corporate engagement from IT-cost centers to top-down strategic partnerships focused on profit-sharing models and tangible business results.

💡 詢問阿法，看詳細逐字稿分析→

Further research

Tickers and signals often linked to this episode's themes in public sources · AI-compiled, not investment advice

Transition to Inference-Driven Compute

Run-time inference is scaling exponentially, shifting focus and capital toward ultra-low latency hardware and localized processing as agentic workloads proliferate.

US stocks

NVDA
NvidiaBenefitsSecured licensing rights to Groq's low-latency LPU architecture and acqui-hired its core leadership team in a $20 billion deal to consolidate its dominance in the real-time inference market.
AMD
AMDBenefitsPositioned to capture high-growth inference and agentic workloads through its high-bandwidth memory Instinct GPU lineup and Venice EPYC processors.
MU
Micron TechnologyBenefitsSupplies the critical high-bandwidth memory required to run massive, distributed inference models at low latencies.

Risks

A deceleration in token consumption or a structural delay in complex reasoning model deployments could leave expensive high-performance inference hardware underutilized.

Watch list

Inference token pricing trends from leading cloud providers
Quarterly HBM shipment volumes from memory producers
Enterprise adoption rates of reasoning-heavy AI models

Daily Active Agents as a KPI

The shift from per-seat SaaS models to outcome-based AI agent platforms is forcing companies to redesign monetization frameworks around work execution.

US stocks

CRM
SalesforceBenefitsPioneering the agent monetization shift with Agentforce by introducing pay-per-resolution pricing, charging enterprises only when an autonomous help agent successfully resolves a customer issue.
TEAM
AtlassianBenefitsMonetizes its Rovo AI platform through credit-based and outcome-linked billing, such as charging flat fees per autonomous resolution for its customer service agents.
ASAN
AsanaPressuredHighly vulnerable to seat-count contraction as autonomous AI agents execute multi-step project tracking and coordination tasks, reducing the need for manual, seat-based licenses.

Risks

Complex legal and attribution issues over what defines a 'resolved outcome' could delay corporate adoption of pay-for-results pricing models.

Watch list

Active agent tracking metrics in quarterly SaaS earnings calls
Gross revenue retention rates at seat-based software vendors
Public pricing adjustments for enterprise AI copilots

Silicon Vertical Integration

Hyperscalers and AI giants are designing proprietary custom chips to bypass GPU supply bottlenecks, reduce energy consumption, and optimize performance for specific workloads.

US stocks

AVGO
BroadcomBenefitsFunctions as the primary custom silicon co-designer for leading tech giants, recently partnering with OpenAI to design their first in-house inference chip, 'Jalapeño'.
MRVL
Marvell TechnologyBenefitsLeverages its rich advanced-node IP to co-design custom AI ASICs for major cloud operators like Amazon and Microsoft.
TSM
Taiwan SemiconductorBenefitsActs as the exclusive high-volume manufacturing partner fabricating custom hyperscaler silicon on advanced 3nm and A14 process nodes.
NVDA
NvidiaPressuredFaces pressure as custom proprietary ASICs deployed by cloud giants reduce their reliance on third-party merchant GPUs and threaten high hardware margins.

Risks

The high upfront capital and engineering costs of bespoke silicon design could yield sub-optimal performance compared to rapidly advancing general-purpose accelerators.

Watch list

Custom ASIC server shipment market share figures
Quarterly design-pipeline backlog updates from Broadcom and Marvell
Capital expenditure allocation changes among the major cloud hyperscalers

This section is AI-compiled from public sources, may be inaccurate or outdated, is for research reference only, and is not investment advice.

Want the full episode?